Akamai adds AI inference and an agent sandbox

Akamai Inference gives teams managed access to open-weight and custom models through OpenAI-compatible APIs, so existing applications move without rewrites, according to Akamai. Customers choose serverless, pay-per-use inference or dedicated deployments. At launch, Akamai says, the catalog includes 12 open-weight models ranging from 8 billion to 2.8 trillion parameters, along with embedding models for retrieval-augmented generation, and Akamai Cloud's managed databases and object storage let models draw on a company's own data. Akamai's beta documentation lists 12 models in total, one of them an embedding model (Qwen3-Embedding-8B, the 8-billion low end; Moonshot AI's Kimi K3 is the 2.8-trillion top). It names NVIDIA Nemotron 3 Ultra, OpenAI GPT OSS 120B, and Gemma 4 26B A4B IT, and lists the rest from DeepSeek, Z.ai, and Qwen. “Open-weight” is Akamai's description; this check did not review each model's licence.

Akamai says it has been building on the view that AI inference will account for the majority of enterprise focus and investment. Chief Operating Officer Adam Karon said customers want to build, run, and protect AI on one platform: “They want to build, run, and protect AI on the same platform. Akamai Inference and Akamai Agent Sandbox add the inference and execution layers that turn our AI infrastructure into a full AI stack.” Akamai's release also says its AI infrastructure has secured multiyear Akamai Cloud contracts collectively worth more than $14 billion; that is the company's own figure for its broader cloud business, not revenue from these two products.

The sandbox is aimed at agents that write and run their own code; Akamai says that code often executes seconds after it is generated, before anyone reads it. Akamai Agent Sandbox runs the agent on Akamai Functions and puts the code the agent writes in its own isolated environment, sequestered from the agent's runtime. Isolation comes from WebAssembly, delivered as an open-source component of the Spin framework, with limits on compute, memory, outbound connections, and system capabilities.

Akamai's product page describes two layers: the agent runs in a WebAssembly sandbox as a Spin application on Akamai Functions, and the JavaScript or Python code it writes runs in a restricted runtime nested inside that sandbox, with no outbound access by default. Akamai says limits are enforced by the runtime rather than by the agent. Developers can build locally with the open-source Spin SDK and deploy the same agent to Akamai Functions. The current tech preview supports intermittent agents, with each step completing in about a minute before the agent stops until its next instruction; Akamai says stateful components will add support for agents that stay alive for days or weeks.

Akamai also expanded its GPU fleet. Each eight-GPU Blackwell Ultra node carries 288 GB of HBM3e memory per GPU, 2.3 TB in total; Akamai says the architecture targets test-time scaling, chain-of-thought reasoning, and long-context agentic work. The fleet also includes RTX PRO 6000 Blackwell Server Edition GPUs for low-latency inference, and Akamai says it is preparing its data centers for NVIDIA's Vera Rubin platform in 2027. The network underneath spans more than 4,350 edge locations, which is Akamai's argument for running inference close to users and their data.

What the press release does not include: pricing for either product or a general-availability date. Akamai's Inference beta documentation says usage is not billed during the beta, accounts have a monthly token limit, models accept and return text only, and prompts and outputs are not retained (zero data retention). It says the Model Library in Cloud Manager shows per-token prices; this check could not view that library, so no prices are quoted here. The release says the Inference beta includes serverless deployment options and directs teams to contact Akamai Sales about NVIDIA Blackwell Ultra (B300) cluster availability. Open items: whether the model catalog grows, what pricing applies after the beta, and whether teams using the Spin SDK publish results from running real agents in the sandbox.